Skip to content

Latest commit

 

History

History
174 lines (126 loc) · 6.37 KB

1.x.md

File metadata and controls

174 lines (126 loc) · 6.37 KB

Vitis AI Release Notes

Release 1.1

New Features

Model Zoo

  • Model quantization accuracy update
  • Model test and retraining improved
  • Caffe_Xilinx updated to version 1.1
  • U50, U200, U250 performance added

Quantizer

  • Add Tensorflow 1.15 support
  • Bugfixes

Compiler

  • Support cross compilation for Zynq and ZU+ based platforms
  • Vitis AI Compiler for U50
    • Based on the new XIR (Xilinx Intermediate Representation)
    • Support DPUv3E
    • Tested with 40 models from Vitis AI Model Zoo
  • Vitis AI Compiler for Zynq and ZU+
    • Support DPUv2 1.4.1 instruction set
    • Support bias rightward-shift computation to improve model accuracy
    • Support bilinear upsampling operator

Runtime

  • VART (Vitis AI Runtime)
    • Unified runtime based on XIR for Zynq, ZU+ and Alveo
    • Include new APIs for NN performance improvement
    • 7 samples with VART APIs provided

DPU

Vitis AI Library

  • All source code open source
  • Support VART
  • Add support for Alveo
  • Support batch model for DPUv3E

Example & Demo

  • Whole Application Acceleration Example
  • End-to-end pipeline which includes JPEG decode, Resize, CNN inference on Alveo
  • Neptune demo: Use FPGA for multi-stream and multi-model mode
  • AKS demo: Building complex application using C++ and threads

Others

  • TVM (Early access, provide docker upon request)
    • Supported frontends: TFLite, ONNX, MxNet and Pytorch
    • Platform support: ZCU102, ZC104, U200 and U250
    • Tested for 15 models including classification, detection and segmentation from various frameworks
  • xButler upgraded to version 3.0 and provides support for docker container.
  • Improved support on upsampling, deconvolution and large convolutions for segmentation models including FPN for DPUv1

Known Issues

  • Alveo U50 toolchain doesn't support Conv2DTranspose trained in Keras and converted to TF 1.15, which will be fixed in Vitis AI 1.2 release.

Updates

  • 5/6/20 - Fixed hardware bug which will lead to computation errors in some corner case for Alveo U50 Production shell xclbin.
  • 5/6/20 - Added support for Alveo U50 using EA x4 shell for increased performance.


Release 1.0

New Features

Model Zoo

  • Release custom Caffe framework distribution caffe_xilinx
  • Add accuracy test code and retrain code for all Caffe models
  • Increase Tensorflow models to 19 with float/fixed model versions and accuracy test code, including popular models such as SSD, YOLOv3, MLPerf:ssd_resnet34, etc.
  • Add multi-task Caffe model for ADAS applications

Optimizer (A separate package which requires licensing)

  • Caffe Pruning
    • Support for depthwise convolution layer
    • Remove internal implementation-related parameters in transformed prototxt
  • TensorFlow Pruning
    • Release pruning tool based on TensorFlow 1.12
    • Add more validations to user-specified parameters
    • Bug fixes for supporting more networks
  • Darknet pruning
    • new interface for pruning tool
    • support yolov3-spp

Quantizer

  • Tensorflow quantization

    • Support DPU simulation and dumping quantize simulation results.
    • Improve support for some layers and node patterns, including tf.keras.layers.Conv2DTranspose, tf.keras.Dense, tf.keras.layers.LeakyReLU, tf.conv2d + tf.mul
    • Move temp quantize info files from /tmp/ to $output_dir/temp folder, to support multi-users on one machine
    • Bugfixes

  • Caffe quantization

    • Enhanced activation data dump function
    • Ubuntu 18 support
    • Non-unified bit width quantization support
    • Support HDF5 data layer
    • Support of scale layers without parameters but with multiple inputs

Compiler

  • Support cross compilation for Zynq and ZU+ based platforms
  • Enhancements and bug fixes for a broader set of Tensorflow models
  • New Split IO memory model enablement for performance optimization
  • Improved code generation
  • Support Caffe/TensorFlow model compilation over cloud DPU V3E (Early Access)

Runtime

  • Enable edge to cloud deployment over XRT 2019.2
  • Offer the unified Vitis AI C++/Python programming APIs
  • DPU priority-based scheduling and DPU core affinity
  • Introduce adaptive operating layer to unify runtime’s underlying interface for Linux, XRT and QNX
  • QNX RTOS enablement to support automotive customers.
  • Neptune API for X+ML
  • Performance improvements

DPU

  • DPUv2 for Zynq and ZU+

    • Support Vitis flow with reference design based on ZCU102
    • The same DPU also supports Vivado flow
    • All features are configurable
    • Fixed several bugs

  • DPUv3 for U50/U280 (Early access)

Vitis AI Library

  • Support of new Vitis AI Runtime - Vitis AI Library is updated to be based on the new Vitis AI Runtime with unified APIs. It also fully supports XRT 2019.2.
  • New DPU support - Besides DPUv2 for Zynq and ZU+, a new AI Library will support new DPUv3 IPs for Alveo/Cloud using same codes (Early access).
  • New Tensorflow model support - There are up to 19 tensorflow models supported, which are from official tensorflow repository
  • New libraries and demos - There are two new libraries “libdpmultitask” and “libdptfssd” which supports multi-task models and Tensorflow SSD models. An updated classification demo is included to shows how to uses unified APIs in Vitis AI runtime.
  • New Open Source Library - The “libdpbase” library is open source in this release, which shows how to use unified APIs in Vitis AI runtime to construct high-level libraries.
  • New Installation Method - The host side environment adopts uses image installation, which simplifies and unifies the installation process.

Others

  • Support for TVM which enables support for Pytorch, ONNX and SageMaker NEO
  • Partitioning of Tensorflow models and support for xDNNv3 execution in Tensorflow natively
  • Automated Tensorflow model partition, compilation and deployment over DPUv3 (Early access)
  • Butler API for following:
    • Automatic resource discovery and management
    • Multiprocess support – Ability for many containers/processes to access single FPGA
    • FPGA slicing – Ability to use part of FPGA
    • Scaleout support for multiple FPGA on same server
  • Support for pix2pix models

Known Issues